| blob | b5327b4cdb34aa34b9714ceb1d149412e02877be |
1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * alloc.c
5 *
6 * Extent allocs and frees
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/swap.h>
32 #define MLOG_MASK_PREFIX ML_DISK_ALLOC
33 #include <cluster/masklog.h>
56 /*
57 * Structures which describe a path through a btree, and functions to
58 * manipulate them.
59 *
60 * The idea here is to be as generic as possible with the tree
61 * manipulation code.
62 */
66 };
68 #define OCFS2_MAX_PATH_DEPTH 5
73 };
75 #define path_root_bh(_path) ((_path)->p_node[0].bh)
76 #define path_root_el(_path) ((_path)->p_node[0].el)
77 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
78 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
79 #define path_num_items(_path) ((_path)->p_tree_depth + 1)
81 /*
82 * Reset the actual path elements so that we can re-use the structure
83 * to build another path. Generally, this involves freeing the buffer
84 * heads.
85 */
87 {
100 }
102 /*
103 * Tree depth may change during truncate, or insert. If we're
104 * keeping the root extent list, then make sure that our path
105 * structure reflects the proper depth.
106 */
111 }
114 {
118 }
119 }
121 /*
122 * Make the *dest path the same as src and re-initialize src path to
123 * have a root only.
124 */
126 {
139 }
140 }
142 /*
143 * Insert an extent block at given index.
144 *
145 * This will not take an additional reference on eb_bh.
146 */
149 {
152 /*
153 * Right now, no root bh is an extent block, so this helps
154 * catch code errors with dinode trees. The assertion can be
155 * safely removed if we ever need to insert extent block
156 * structures at the root.
157 */
162 }
166 {
177 }
180 }
182 /*
183 * Allocate and initialize a new path based on a disk inode tree.
184 */
186 {
191 }
193 /*
194 * Convenience function to journal all components in a path.
195 */
198 {
206 OCFS2_JOURNAL_ACCESS_WRITE);
210 }
211 }
213 out:
215 }
219 CONTIG_LEFT,
220 CONTIG_RIGHT
221 };
224 /*
225 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
226 * ocfs2_extent_contig only work properly against leaf nodes!
227 */
230 u64 blkno)
231 {
238 }
242 {
243 u32 left_range;
249 }
251 static enum ocfs2_contig_type
255 {
268 }
270 /*
271 * NOTE: We can have pretty much any combination of contiguousness and
272 * appending.
273 *
274 * The usefulness of APPEND_TAIL is more in that it lets us know that
275 * we'll have to update the path to that leaf.
276 */
279 APPEND_TAIL,
280 };
288 };
290 /*
291 * How many free extents have we got before we need more meta data?
292 */
296 {
308 }
316 }
325 bail:
331 }
333 /* expects array to already be allocated
334 *
335 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
336 * l_count for you
337 */
344 {
346 u16 suballoc_bit_start;
347 u32 num_got;
348 u64 first_blkno;
356 handle,
357 meta_ac,
365 }
373 }
377 OCFS2_JOURNAL_ACCESS_CREATE);
381 }
385 /* Ok, setup the minimal stuff here. */
394 suballoc_bit_start++;
395 first_blkno++;
397 /* We'll also be dirtied by the caller, so
398 * this isn't absolutely necessary. */
403 }
404 }
407 }
410 bail:
416 }
417 }
420 }
422 /*
423 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
424 *
425 * Returns the sum of the rightmost extent rec logical offset and
426 * cluster count.
427 *
428 * ocfs2_add_branch() uses this to determine what logical cluster
429 * value should be populated into the leftmost new branch records.
430 *
431 * ocfs2_shift_tree_depth() uses this to determine the # clusters
432 * value for the new topmost tree record.
433 */
435 {
442 }
444 /*
445 * Add an entire tree branch to our inode. eb_bh is the extent block
446 * to start at, if we don't want to start the branch at the dinode
447 * structure.
448 *
449 * last_eb_bh is required as we have to update it's next_leaf pointer
450 * for the new last extent block.
451 *
452 * the new branch will be 'empty' in the sense that every block will
453 * contain a single record with cluster count == 0.
454 */
462 {
471 u32 new_cpos;
485 /* we never add a branch to a leaf. */
490 /* allocate the number of new eb blocks we need */
492 GFP_KERNEL);
497 }
504 }
509 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
510 * linked with the rest of the tree.
511 * conversly, new_eb_bhs[0] is the new bottommost leaf.
512 *
513 * when we leave the loop, new_last_eb_blk will point to the
514 * newest leaf, and next_blkno will point to the topmost extent
515 * block. */
524 }
528 OCFS2_JOURNAL_ACCESS_CREATE);
532 }
537 /*
538 * This actually counts as an empty extent as
539 * c_clusters == 0
540 */
543 /*
544 * eb_el isn't always an interior node, but even leaf
545 * nodes want a zero'd flags and reserved field so
546 * this gets the whole 32 bits regardless of use.
547 */
556 }
559 }
561 /* This is a bit hairy. We want to update up to three blocks
562 * here without leaving any of them in an inconsistent state
563 * in case of error. We don't have to worry about
564 * journal_dirty erroring as it won't unless we've aborted the
565 * handle (in which case we would never be here) so reserving
566 * the write with journal_access is all we need to do. */
568 OCFS2_JOURNAL_ACCESS_WRITE);
572 }
574 OCFS2_JOURNAL_ACCESS_WRITE);
578 }
581 OCFS2_JOURNAL_ACCESS_WRITE);
585 }
586 }
588 /* Link the new branch into the rest of the tree (el will
589 * either be on the fe, or the extent block passed in. */
596 /* fe needs a new last extent block pointer, as does the
597 * next_leaf on the previously last-extent-block. */
613 }
616 bail:
622 }
626 }
628 /*
629 * adds another level to the allocation tree.
630 * returns back the new extent block so you can add a branch to it
631 * after this call.
632 */
639 {
641 u32 new_clusters;
655 }
662 }
669 OCFS2_JOURNAL_ACCESS_CREATE);
673 }
675 /* copy the fe data into the new extent block */
685 }
688 OCFS2_JOURNAL_ACCESS_WRITE);
692 }
696 /* update fe now */
705 /* If this is our 1st tree depth shift, then last_eb_blk
706 * becomes the allocated extent block */
714 }
719 bail:
725 }
727 /*
728 * Should only be called when there is no space left in any of the
729 * leaf nodes. What we want to do is find the lowest tree depth
730 * non-leaf extent block with room for new records. There are three
731 * valid results of this search:
732 *
733 * 1) a lowest extent block is found, then we pass it back in
734 * *lowest_eb_bh and return '0'
735 *
736 * 2) the search fails to find anything, but the dinode has room. We
737 * pass NULL back in *lowest_eb_bh, but still return '0'
738 *
739 * 3) the search fails to find anything AND the dinode is full, in
740 * which case we return > 0
741 *
742 * return status < 0 indicates an error.
743 */
748 {
750 u64 blkno;
771 }
776 "list where extent # %d has no physical "
781 }
786 }
789 inode);
793 }
800 }
809 }
810 }
812 /* If we didn't find one and the fe doesn't have any room,
813 * then return '1' */
819 bail:
825 }
827 /*
828 * This is only valid for leaf nodes, which are the only ones that can
829 * have empty extents anyway.
830 */
832 {
834 }
836 /*
837 * This function will discard the rightmost extent record.
838 */
840 {
846 /* This will cause us to go off the end of our extent list. */
852 }
856 {
866 /* The tree code before us didn't allow enough room in the leaf. */
870 /*
871 * The easiest way to approach this is to just remove the
872 * empty extent and temporarily decrement next_free.
873 */
875 /*
876 * If next_free was 1 (only an empty extent), this
877 * loop won't execute, which is fine. We still want
878 * the decrement above to happen.
879 */
883 next_free--;
884 }
886 /*
887 * Figure out what the new record index should be.
888 */
894 }
904 /*
905 * No need to memmove if we're just adding to the tail.
906 */
914 num_bytes);
915 }
917 /*
918 * Either we had an empty extent, and need to re-increment or
919 * there was no empty extent on a non full rightmost leaf node,
920 * in which case we still need to increment.
921 */
922 next_free++;
924 /*
925 * Make sure none of the math above just messed up our tree.
926 */
931 }
933 /*
934 * Create an empty extent record .
935 *
936 * l_next_free_rec may be updated.
937 *
938 * If an empty extent already exists do nothing.
939 */
941 {
960 set_and_inc:
963 }
965 /*
966 * For a rotation which involves two leaf nodes, the "root node" is
967 * the lowest level tree node which contains a path to both leafs. This
968 * resulting set of information can be used to form a complete "subtree"
969 *
970 * This function is passed two full paths from the dinode down to a
971 * pair of adjacent leaves. It's task is to figure out which path
972 * index contains the subtree root - this can be the root index itself
973 * in a worst-case rotation.
974 *
975 * The array index of the subtree root is passed back.
976 */
980 {
983 /*
984 * Check that the caller passed in two paths from the same tree.
985 */
989 i++;
991 /*
992 * The caller didn't pass two adjacent paths.
993 */
1005 }
1009 /*
1010 * Traverse a btree path in search of cpos, starting at root_el.
1011 *
1012 * This code can be called with a cpos larger than the tree, in which
1013 * case it will return the rightmost path.
1014 */
1018 {
1020 u32 range;
1021 u64 blkno;
1032 "Inode %llu has empty extent list at "
1039 }
1044 /*
1045 * In the case that cpos is off the allocation
1046 * tree, this should just wind up returning the
1047 * rightmost record.
1048 */
1053 }
1058 "Inode %llu has bad blkno in extent list "
1064 }
1073 }
1081 }
1086 "Inode %llu has bad count in extent list "
1094 }
1098 }
1100 out:
1101 /*
1102 * Catch any trailing bh that the loop didn't handle.
1103 */
1107 }
1109 /*
1110 * Given an initialized path (that is, it has a valid root extent
1111 * list), this function will traverse the btree in search of the path
1112 * which would contain cpos.
1113 *
1114 * The path traveled is recorded in the path structure.
1115 *
1116 * Note that this will not do any comparisons on leaf node extent
1117 * records, so it will work fine in the case that we just added a tree
1118 * branch.
1119 */
1123 };
1125 {
1131 }
1133 u32 cpos)
1134 {
1141 }
1144 {
1149 /* We want to retain only the leaf block. */
1153 }
1154 }
1155 /*
1156 * Find the leaf block in the tree which would contain cpos. No
1157 * checking of the actual leaf is done.
1158 *
1159 * Some paths want to call this instead of allocating a path structure
1160 * and calling ocfs2_find_path().
1161 *
1162 * This function doesn't handle non btree extent lists.
1163 */
1166 {
1174 }
1177 out:
1179 }
1181 /*
1182 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
1183 *
1184 * Basically, we've moved stuff around at the bottom of the tree and
1185 * we need to fix up the extent records above the changes to reflect
1186 * the new changes.
1187 *
1188 * left_rec: the record on the left.
1189 * left_child_el: is the child list pointed to by left_rec
1190 * right_rec: the record to the right of left_rec
1191 * right_child_el: is the child list pointed to by right_rec
1192 *
1193 * By definition, this only works on interior nodes.
1194 */
1199 {
1202 /*
1203 * Interior nodes never have holes. Their cpos is the cpos of
1204 * the leftmost record in their child list. Their cluster
1205 * count covers the full theoretical range of their child list
1206 * - the range between their cpos and the cpos of the record
1207 * immediately to their right.
1208 */
1213 /*
1214 * Calculate the rightmost cluster count boundary before
1215 * moving cpos - we will need to adjust clusters after
1216 * updating e_cpos to keep the same highest cluster count.
1217 */
1226 }
1228 /*
1229 * Adjust the adjacent root node records involved in a
1230 * rotation. left_el_blkno is passed in as a key so that we can easily
1231 * find it's index in the root list.
1232 */
1236 u64 left_el_blkno)
1237 {
1246 }
1248 /*
1249 * The path walking code should have never returned a root and
1250 * two paths which are not adjacent.
1251 */
1256 }
1258 /*
1259 * We've changed a leaf block (in right_path) and need to reflect that
1260 * change back up the subtree.
1261 *
1262 * This happens in multiple places:
1263 * - When we've moved an extent record from the left path leaf to the right
1264 * path leaf to make room for an empty extent in the left path leaf.
1265 * - When our insert into the right path leaf is at the leftmost edge
1266 * and requires an update of the path immediately to it's left. This
1267 * can occur at the end of some types of rotation and appending inserts.
1268 */
1273 {
1279 /*
1280 * Update the counts and position values within all the
1281 * interior nodes to reflect the leaf rotation we just did.
1282 *
1283 * The root node is handled below the loop.
1284 *
1285 * We begin the loop with right_el and left_el pointing to the
1286 * leaf lists and work our way up.
1287 *
1288 * NOTE: within this loop, left_el and right_el always refer
1289 * to the *child* lists.
1290 */
1296 /*
1297 * One nice property of knowing that all of these
1298 * nodes are below the root is that we only deal with
1299 * the leftmost right node record and the rightmost
1300 * left node record.
1301 */
1310 right_el);
1320 /*
1321 * Setup our list pointers now so that the current
1322 * parents become children in the next iteration.
1323 */
1326 }
1328 /*
1329 * At the root node, adjust the two adjacent records which
1330 * begin our path to the leaves.
1331 */
1345 }
1352 {
1365 "Inode %llu has non-full interior leaf node %llu"
1371 }
1373 /*
1374 * This extent block may already have an empty record, so we
1375 * return early if so.
1376 */
1384 OCFS2_JOURNAL_ACCESS_WRITE);
1388 }
1393 OCFS2_JOURNAL_ACCESS_WRITE);
1397 }
1401 OCFS2_JOURNAL_ACCESS_WRITE);
1405 }
1406 }
1411 /* This is a code error, not a disk corruption. */
1423 }
1425 /* Do the copy now. */
1430 /*
1431 * Clear out the record we just copied and shift everything
1432 * over, leaving an empty extent in the left leaf.
1433 *
1434 * We temporarily subtract from next_free_rec so that the
1435 * shift will lose the tail record (which is now defunct).
1436 */
1446 }
1449 subtree_index);
1451 out:
1453 }
1455 /*
1456 * Given a full path, determine what cpos value would return us a path
1457 * containing the leaf immediately to the left of the current one.
1458 *
1459 * Will return zero if the path passed in is already the leftmost path.
1460 */
1463 {
1465 u64 blkno;
1474 /* Start at the tree node just above the leaf and work our way up. */
1479 /*
1480 * Find the extent record just before the one in our
1481 * path.
1482 */
1487 /*
1488 * We've determined that the
1489 * path specified is already
1490 * the leftmost one - return a
1491 * cpos of zero.
1492 */
1494 }
1495 /*
1496 * The leftmost record points to our
1497 * leaf - we need to travel up the
1498 * tree one level.
1499 */
1501 }
1508 }
1509 }
1511 /*
1512 * If we got here, we never found a valid node where
1513 * the tree indicated one should be.
1514 */
1521 next_node:
1523 i--;
1524 }
1526 out:
1528 }
1532 {
1539 }
1541 /*
1542 * Trap the case where we're inserting into the theoretical range past
1543 * the _actual_ left leaf range. Otherwise, we'll rotate a record
1544 * whose cpos is less than ours into the right leaf.
1545 *
1546 * It's only necessary to look at the rightmost record of the left
1547 * leaf because the logic that calls us should ensure that the
1548 * theoretical ranges in the path components above the leaves are
1549 * correct.
1550 */
1552 u32 insert_cpos)
1553 {
1565 }
1567 /*
1568 * Rotate all the records in a btree right one record, starting at insert_cpos.
1569 *
1570 * The path to the rightmost leaf should be passed in.
1571 *
1572 * The array is assumed to be large enough to hold an entire path (tree depth).
1573 *
1574 * Upon succesful return from this function:
1575 *
1576 * - The 'right_path' array will contain a path to the leaf block
1577 * whose range contains e_cpos.
1578 * - That leaf block will have a single empty extent in list index 0.
1579 * - In the case that the rotation requires a post-insert update,
1580 * *ret_left_path will contain a valid path which can be passed to
1581 * ocfs2_insert_path().
1582 */
1585 u32 insert_cpos,
1588 {
1590 u32 cpos;
1601 }
1607 }
1611 /*
1612 * What we want to do here is:
1613 *
1614 * 1) Start with the rightmost path.
1615 *
1616 * 2) Determine a path to the leaf block directly to the left
1617 * of that leaf.
1618 *
1619 * 3) Determine the 'subtree root' - the lowest level tree node
1620 * which contains a path to both leaves.
1621 *
1622 * 4) Rotate the subtree.
1623 *
1624 * 5) Find the next subtree by considering the left path to be
1625 * the new right path.
1626 *
1627 * The check at the top of this while loop also accepts
1628 * insert_cpos == cpos because cpos is only a _theoretical_
1629 * value to get us the left path - insert_cpos might very well
1630 * be filling that hole.
1631 *
1632 * Stop at a cpos of '0' because we either started at the
1633 * leftmost branch (i.e., a tree with one branch and a
1634 * rotation inside of it), or we've gone as far as we can in
1635 * rotating subtrees.
1636 */
1645 }
1649 "Inode %lu: error during insert of %u "
1650 "(left path cpos %u) results in two identical "
1657 insert_cpos)) {
1660 /*
1661 * We've rotated the tree as much as we
1662 * should. The rest is up to
1663 * ocfs2_insert_path() to complete, after the
1664 * record insertion. We indicate this
1665 * situation by returning the left path.
1666 *
1667 * The reason we don't adjust the records here
1668 * before the record insert is that an error
1669 * later might break the rule where a parent
1670 * record e_cpos will reflect the actual
1671 * e_cpos of the 1st nonempty record of the
1672 * child list.
1673 */
1676 }
1681 start,
1686 right_path);
1690 }
1697 }
1699 /*
1700 * There is no need to re-read the next right path
1701 * as we know that it'll be our current left
1702 * path. Optimize by copying values instead.
1703 */
1711 }
1712 }
1714 out:
1717 out_ret_path:
1719 }
1721 /*
1722 * Do the final bits of extent record insertion at the target leaf
1723 * list. If this leaf is part of an allocation tree, it is assumed
1724 * that the tree above has been prepared.
1725 */
1730 {
1737 /*
1738 * Contiguous insert - either left or right.
1739 */
1745 }
1749 }
1751 /*
1752 * Handle insert into an empty leaf.
1753 */
1760 }
1762 /*
1763 * Appending insert.
1764 */
1774 "inode %lu, depth %u, count %u, next free %u, "
1775 "rec.cpos %u, rec.clusters %u, "
1785 i++;
1789 }
1791 /*
1792 * Ok, we have to rotate.
1793 *
1794 * At this point, it is safe to assume that inserting into an
1795 * empty leaf and appending to a leaf have both been handled
1796 * above.
1797 *
1798 * This leaf needs to have space, either by the empty 1st
1799 * extent record, or by virtue of an l_next_rec < l_count.
1800 */
1802 }
1806 u32 clusters)
1807 {
1812 }
1818 {
1826 /*
1827 * This shouldn't happen for non-trees. The extent rec cluster
1828 * count manipulation below only works for interior nodes.
1829 */
1832 /*
1833 * If our appending insert is at the leftmost edge of a leaf,
1834 * then we might need to update the rightmost records of the
1835 * neighboring path.
1836 */
1841 u32 left_cpos;
1848 }
1852 left_cpos);
1854 /*
1855 * No need to worry if the append is already in the
1856 * leftmost leaf.
1857 */
1865 }
1871 }
1873 /*
1874 * ocfs2_insert_path() will pass the left_path to the
1875 * journal for us.
1876 */
1877 }
1878 }
1884 }
1899 }
1913 /* Don't touch the leaf node */
1919 }
1923 out:
1928 }
1930 /*
1931 * This function only does inserts on an allocation b-tree. For dinode
1932 * lists, ocfs2_insert_at_leaf() is called directly.
1933 *
1934 * right_path is the path we want to do the actual insert
1935 * in. left_path should only be passed in if we need to update that
1936 * portion of the tree after an edge insert.
1937 */
1944 {
1949 /*
1950 * Pass both paths to the journal. The majority of inserts
1951 * will be touching all components anyway.
1952 */
1957 }
1962 /*
1963 * There's a chance that left_path got passed back to
1964 * us without being accounted for in the
1965 * journal. Extend our transaction here to be sure we
1966 * can change those blocks.
1967 */
1974 }
1980 }
1981 }
1991 /*
1992 * The rotate code has indicated that we need to fix
1993 * up portions of the tree after the insert.
1994 *
1995 * XXX: Should we extend the transaction here?
1996 */
1998 right_path);
2001 }
2004 out:
2006 }
2013 {
2015 u32 cpos;
2025 OCFS2_JOURNAL_ACCESS_WRITE);
2029 }
2034 }
2041 }
2043 /*
2044 * Determine the path to start with. Rotations need the
2045 * rightmost path, everything else can go directly to the
2046 * target leaf.
2047 */
2053 }
2059 }
2061 /*
2062 * Rotations and appends need special treatment - they modify
2063 * parts of the tree's above them.
2064 *
2065 * Both might pass back a path immediate to the left of the
2066 * one being inserted to. This will be cause
2067 * ocfs2_insert_path() to modify the rightmost records of
2068 * left_path to account for an edge insert.
2069 *
2070 * XXX: When modifying this code, keep in mind that an insert
2071 * can wind up skipping both of these two special cases...
2072 */
2080 }
2088 }
2089 }
2096 }
2098 out_update_clusters:
2106 out:
2111 }
2117 {
2125 insert_rec);
2129 }
2130 }
2132 }
2134 /*
2135 * This should only be called against the righmost leaf extent list.
2136 *
2137 * ocfs2_figure_appending_type() will figure out whether we'll have to
2138 * insert at the tail of the rightmost leaf.
2139 *
2140 * This should also work against the dinode list for tree's with 0
2141 * depth. If we consider the dinode list to be the rightmost leaf node
2142 * then the logic here makes sense.
2143 */
2147 {
2160 /* Were all records empty? */
2163 }
2174 set_tail_append:
2176 }
2178 /*
2179 * Helper function called at the begining of an insert.
2180 *
2181 * This computes a few things that are commonly used in the process of
2182 * inserting into the btree:
2183 * - Whether the new extent is contiguous with an existing one.
2184 * - The current tree depth.
2185 * - Whether the insert is an appending one.
2186 * - The total # of free records in the tree.
2187 *
2188 * All of the information is stored on the ocfs2_insert_type
2189 * structure.
2190 */
2196 {
2208 /*
2209 * If we have tree depth, we read in the
2210 * rightmost extent block ahead of time as
2211 * ocfs2_figure_insert_type() and ocfs2_add_branch()
2212 * may want it later.
2213 */
2220 }
2223 }
2225 /*
2226 * Unless we have a contiguous insert, we'll need to know if
2227 * there is room left in our allocation tree for another
2228 * extent record.
2229 *
2230 * XXX: This test is simplistic, we can search for empty
2231 * extent records too.
2232 */
2240 }
2247 }
2249 /*
2250 * In the case that we're inserting past what the tree
2251 * currently accounts for, ocfs2_find_path() will return for
2252 * us the rightmost tree path. This is accounted for below in
2253 * the appending code.
2254 */
2259 }
2263 /*
2264 * Now that we have the path, there's two things we want to determine:
2265 * 1) Contiguousness (also set contig_index if this is so)
2266 *
2267 * 2) Are we doing an append? We can trivially break this up
2268 * into two types of appends: simple record append, or a
2269 * rotate inside the tail leaf.
2270 */
2273 /*
2274 * The insert code isn't quite ready to deal with all cases of
2275 * left contiguousness. Specifically, if it's an insert into
2276 * the 1st record in a leaf, it will require the adjustment of
2277 * cluster count on the last record of the path directly to it's
2278 * left. For now, just catch that case and fool the layers
2279 * above us. This works just fine for tree_depth == 0, which
2280 * is why we allow that above.
2281 */
2286 /*
2287 * Ok, so we can simply compare against last_eb to figure out
2288 * whether the path doesn't exist. This will only happen in
2289 * the case that we're doing a tail append, so maybe we can
2290 * take advantage of that information somehow.
2291 */
2293 /*
2294 * Ok, ocfs2_find_path() returned us the rightmost
2295 * tree path. This might be an appending insert. There are
2296 * two cases:
2297 * 1) We're doing a true append at the tail:
2298 * -This might even be off the end of the leaf
2299 * 2) We're "appending" by rotating in the tail
2300 */
2302 }
2304 out:
2309 else
2312 }
2314 /*
2315 * Insert an extent into an inode btree.
2316 *
2317 * The caller needs to update fe->i_clusters
2318 */
2323 u32 cpos,
2324 u64 start_blk,
2325 u32 new_clusters,
2327 {
2339 "Device %s, asking for sparse allocation: inode %llu, "
2355 }
2358 "Insert.contig_index: %d, Insert.free_records: %d, "
2363 /*
2364 * Avoid growing the tree unless we're out of records and the
2365 * insert type requres one.
2366 */
2375 }
2377 /* We traveled all the way to the bottom of the allocation tree
2378 * and didn't find room for any more extents - we need to add
2379 * another tree level */
2385 /* ocfs2_shift_tree_depth will return us a buffer with
2386 * the new extent block (so we can pass that to
2387 * ocfs2_add_branch). */
2393 }
2395 /* Special case: we have room now if we shifted from
2396 * tree_depth 0 */
2399 }
2401 /* call ocfs2_add_branch to add the final part of the tree with
2402 * the new data. */
2405 meta_ac);
2409 }
2411 out_add:
2412 /* Finally, we can add clusters. This might rotate the tree for us. */
2416 else
2419 bail:
2428 }
2431 {
2440 "slot %d, invalid truncate log parameters: used = "
2444 }
2448 {
2452 /* No records, nothing to coalesce */
2461 }
2465 u64 start_blk,
2467 {
2488 }
2493 "Truncate record count on #%llu invalid "
2499 /* Caller should have known to flush before calling us. */
2505 }
2508 OCFS2_JOURNAL_ACCESS_WRITE);
2512 }
2519 /*
2520 * Move index back to the record we are coalescing with.
2521 * ocfs2_truncate_log_can_coalesce() guarantees nonzero
2522 */
2523 index--;
2528 num_clusters);
2532 }
2539 }
2541 bail:
2544 }
2550 {
2554 u64 start_blk;
2567 /* Caller has given us at least enough credits to
2568 * update the truncate log dinode */
2570 OCFS2_JOURNAL_ACCESS_WRITE);
2574 }
2582 }
2584 /* TODO: Perhaps we can calculate the bulk of the
2585 * credits up front rather than extending like
2586 * this. */
2588 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC);
2592 }
2599 /* if start_blk is not set, we ignore the record as
2600 * invalid. */
2607 num_clusters);
2611 }
2612 }
2613 i--;
2614 }
2616 bail:
2619 }
2621 /* Expects you to already be holding tl_inode->i_mutex */
2623 {
2644 }
2652 }
2655 GLOBAL_BITMAP_SYSTEM_INODE,
2656 OCFS2_INVALID_SLOT);
2661 }
2669 }
2676 }
2679 data_alloc_bh);
2685 out_unlock:
2689 out_mutex:
2693 out:
2696 }
2699 {
2708 }
2711 {
2724 }
2726 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ)
2729 {
2731 /* We want to push off log flushes while truncates are
2732 * still running. */
2737 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL);
2738 }
2739 }
2745 {
2751 TRUNCATE_LOG_SYSTEM_INODE,
2752 slot_num);
2757 }
2765 }
2769 bail:
2772 }
2774 /* called during the 1st stage of node recovery. we stamp a clean
2775 * truncate log and pass back a copy for processing later. if the
2776 * truncate log does not require processing, a *tl_copy is set to
2777 * NULL. */
2781 {
2796 }
2804 }
2815 }
2817 /* Assuming the write-out below goes well, this copy
2818 * will be passed back to recovery for processing. */
2821 /* All we need to do to clear the truncate log is set
2822 * tl_used. */
2829 }
2830 }
2832 bail:
2841 }
2845 }
2849 {
2853 u64 start_blk;
2863 }
2877 }
2878 }
2885 }
2897 }
2898 }
2900 bail_up:
2905 }
2908 {
2924 }
2927 }
2930 {
2944 /* ocfs2_truncate_log_shutdown keys on the existence of
2945 * osb->osb_tl_inode so we don't set any of the osb variables
2946 * until we're sure all is well. */
2948 ocfs2_truncate_log_worker);
2954 }
2956 /*
2957 * Delayed de-allocation of suballocator blocks.
2958 *
2959 * Some sets of block de-allocations might involve multiple suballocator inodes.
2960 *
2961 * The locking for this can get extremely complicated, especially when
2962 * the suballocator inodes to delete from aren't known until deep
2963 * within an unrelated codepath.
2964 *
2965 * ocfs2_extent_block structures are a good example of this - an inode
2966 * btree could have been grown by any number of nodes each allocating
2967 * out of their own suballoc inode.
2968 *
2969 * These structures allow the delay of block de-allocation until a
2970 * later time, when locking of multiple cluster inodes won't cause
2971 * deadlock.
2972 */
2974 /*
2975 * Describes a single block free from a suballocator
2976 */
2979 u64 free_blk;
2981 };
2988 };
2994 {
2996 u64 bg_blkno;
3007 }
3015 }
3022 }
3035 }
3041 }
3046 }
3048 out_journal:
3051 out_unlock:
3054 out_mutex:
3057 out:
3059 /* Premature exit may have left some dangling items. */
3063 }
3066 }
3070 {
3089 }
3093 }
3096 }
3102 {
3110 }
3120 }
3122 }
3127 {
3137 }
3144 }
3156 out:
3158 }
3162 {
3167 }
3169 /* This function will figure out whether the currently last extent
3170 * block will be deleted, and if it will, what the new last extent
3171 * block will be so we can update his h_next_leaf_blk field, as well
3172 * as the dinodes i_last_eb_blk */
3177 {
3179 u32 cpos;
3187 /* we have no tree, so of course, no last_eb. */
3191 /* trunc to zero special case - this makes tree_depth = 0
3192 * regardless of what it is. */
3199 /*
3200 * Make sure that this extent list will actually be empty
3201 * after we clear away the data. We can shortcut out if
3202 * there's more than one non-empty extent in the
3203 * list. Otherwise, a check of the remaining extent is
3204 * necessary.
3205 */
3212 /* We may have a valid extent in index 1, check it. */
3216 /*
3217 * Fall through - no more nonempty extents, so we want
3218 * to delete this leaf.
3219 */
3225 }
3228 /*
3229 * Check it we'll only be trimming off the end of this
3230 * cluster.
3231 */
3234 }
3240 }
3246 }
3254 }
3260 out:
3264 }
3266 /*
3267 * Trim some clusters off the rightmost edge of a tree. Only called
3268 * during truncate.
3269 *
3270 * The caller needs to:
3271 * - start journaling of each path component.
3272 * - compute and fully set up any new last ext block
3273 */
3277 {
3304 }
3306 find_tail_record:
3319 /*
3320 * If the leaf block contains a single empty
3321 * extent and no records, we can just remove
3322 * the block.
3323 */
3330 }
3332 /*
3333 * Remove any empty extents by shifting things
3334 * left. That should make life much easier on
3335 * the code below. This condition is rare
3336 * enough that we shouldn't see a performance
3337 * hit.
3338 */
3349 /*
3350 * We've modified our extent list. The
3351 * simplest way to handle this change
3352 * is to being the search from the
3353 * start again.
3354 */
3356 }
3360 /*
3361 * We'll use "new_edge" on our way back up the
3362 * tree to know what our rightmost cpos is.
3363 */
3367 /*
3368 * The caller will use this to delete data blocks.
3369 */
3374 /*
3375 * If it's now empty, remove this record.
3376 */
3381 }
3389 }
3391 /* Can this actually happen? */
3395 /*
3396 * We never actually deleted any clusters
3397 * because our leaf was empty. There's no
3398 * reason to adjust the rightmost edge then.
3399 */
3407 /*
3408 * A deleted child record should have been
3409 * caught above.
3410 */
3412 }
3419 }
3428 /*
3429 * We must be careful to only attempt delete of an
3430 * extent block (and not the root inode block).
3431 */
3436 /*
3437 * Save this for use when processing the
3438 * parent block.
3439 */
3451 /* An error here is not fatal. */
3456 }
3458 index--;
3459 }
3462 out:
3464 }
3473 {
3488 }
3490 /*
3491 * Each component will be touched, so we might as well journal
3492 * here to avoid having to handle errors later.
3493 */
3498 }
3502 OCFS2_JOURNAL_ACCESS_WRITE);
3506 }
3509 }
3513 /*
3514 * Lower levels depend on this never happening, but it's best
3515 * to check it up here before changing the tree.
3516 */
3523 }
3527 clusters_to_del;
3536 }
3539 /* trunc to zero is a special case. */
3549 }
3552 /* If there will be a new last extent block, then by
3553 * definition, there cannot be any leaves to the right of
3554 * him. */
3560 }
3561 }
3565 clusters_to_del);
3569 }
3570 }
3572 bail:
3576 }
3579 {
3583 }
3586 {
3590 }
3595 {
3609 /*
3610 * Since 'from' has been capped to a value below page
3611 * size, this calculation won't be able to overflow
3612 * 'to'
3613 */
3616 /*
3617 * The truncate tail in this case should never contain
3618 * more than one page at maximum. The loop below also
3619 * assumes this.
3620 */
3622 }
3638 /*
3639 * Need to set the buffers we zero'd into uptodate
3640 * here if they aren't - ocfs2_map_page_blocks()
3641 * might've skipped some
3642 */
3647 ocfs2_ordered_zero_func);
3653 ocfs2_writeback_zero_func);
3656 }
3663 /*
3664 * Every page after the 1st one should be completely zero'd.
3665 */
3667 }
3668 out:
3675 }
3676 }
3677 }
3681 {
3688 u64 next_cluster_bytes;
3692 /* Cluster boundary, so we don't need to grab any pages. */
3701 }
3703 /* Tail is a hole. */
3707 /* Tail is marked as unwritten, we can count on write to zero
3708 * in that case. */
3720 }
3722 numpages++;
3723 index++;
3726 out:
3733 }
3734 }
3735 }
3737 }
3742 }
3744 /*
3745 * Zero the area past i_size but still within an allocated
3746 * cluster. This avoids exposing nonzero data on subsequent file
3747 * extends.
3748 *
3749 * We need to call this before i_size is updated on the inode because
3750 * otherwise block_write_full_page() will skip writeout of pages past
3751 * i_size. The new_i_size parameter is passed for this reason.
3752 */
3754 u64 new_i_size)
3755 {
3757 loff_t endbyte;
3759 u64 phys;
3761 /*
3762 * File systems which don't support sparse files zero on every
3763 * extend.
3764 */
3774 }
3780 }
3786 handle);
3788 /*
3789 * Initiate writeout of the pages we zero'd here. We don't
3790 * wait on them - the truncate_inode_pages() call later will
3791 * do that for us.
3792 */
3799 out:
3804 }
3806 /*
3807 * It is expected, that by the time you call this function,
3808 * inode->i_size and fe->i_size have been adjusted.
3809 *
3810 * WARNING: This will kfree the truncate context
3811 */
3816 {
3834 }
3838 start:
3839 /*
3840 * Check that we still have allocation to delete.
3841 */
3845 }
3847 /*
3848 * Truncate always works against the rightmost tree branch.
3849 */
3854 }
3859 /*
3860 * By now, el will point to the extent list on the bottom most
3861 * portion of this tree. Only the tail record is considered in
3862 * each pass.
3863 *
3864 * We handle the following cases, in order:
3865 * - empty extent: delete the remaining branch
3866 * - remove the entire record
3867 * - remove a partial record
3868 * - no record needs to be removed (truncate has completed)
3869 */
3878 }
3890 new_highest_cpos;
3894 }
3903 /* ocfs2_truncate_log_needs_flush guarantees us at least one
3904 * record is free for use. If there isn't any, we flush to get
3905 * an empty truncate log. */
3911 }
3912 }
3916 el);
3923 }
3930 }
3940 /*
3941 * The check above will catch the case where we've truncated
3942 * away all allocation.
3943 */
3946 bail:
3960 /* This will drop the ext_alloc cluster lock for us */
3965 }
3967 /*
3968 * Expects the inode to already be locked.
3969 */
3974 {
3998 }
4007 }
4015 }
4016 }
4021 bail:
4026 }
4029 }
4032 {
4033 /*
4034 * The caller is responsible for completing deallocation
4035 * before freeing the context.
4036 */
4045 }